NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Hello, everyone, and thank you for joining our webinar, How to create a power-efficient data center with DPUs. In this webinar, you'll be learning why power efficiency is an increasingly important metric for IT and several strategies for improving power efficiency by deploying NVIDIA's BlueField DPUs within your data center. Before we begin, I'd like to cover a few housekeeping items. [Operator Instructions] As well, we have a few poll questions that we'll be asking during the presentation we'd like you to participate by answering to the best of your knowledge. Also, there is a copy of the presentation as well as a few other resources, we believe that will be useful that can be found in the Resource List. We encourage you to download any resources or bookmark links that you may find useful. Next, a little bit about our speaker today, John Kim, who is NVIDIA's Director of Storage Marketing for NVIDIA. John focuses on -- John's focus at NVIDIA is to help customers and vendors benefit from high-performance networking connections through SmartNIC offloads and DPU acceleration. In the areas of storage, Big Data AI and cybersecurity. John frequently speaks at conferences, who may not be a stranger to many of you. I'd also recommend looking at the NVIDIA blog site. You can find more information. John has written some nice blogs, informational blogs on the DPUs. Now without any further ado, I'll turn it over to our speaker, John Kim.

Thank you, Tim, and hello, everyone. So welcome to our webinar. We're going to cover helping it -- how do you build a power-efficient data center using DPUs. So I think the first thing we should talk about is why does energy efficiency matter in the data center? And what are the benefits and how can DPUs help you? And that's what we'll get to later in the Data Center. So we've got 4 big reasons why you should care, why we all should care about power efficiency in the Data Center. So the first is that electricity prices are going up. And that's likely to be a long-term trend, not just a temporary artifact of the comeback from COVID. And then also, there are many data centers where there's a hard limit on how much power you can get in. And that means that no matter how much you're willing to pay, you just cannot get any more electricity into that Data Center. So the only way to have more servers is to make it more efficient. Third, there was a big corporate push, and for government organizations and even nonprofits. There's a big push to make data centers more green. This is part of the ESG or Environmental, Social and Governance movement. So there's a lot of interest in -- from Boards of Directors, from CEOs and from government agencies and finding out how to make data centers more sustainable and the power sourcing greener in those data centers. And finally, there's always pricing pressure from cloud service providers who tend to be very power efficient already and are baking those efficiencies into their pricing. So there is always that pressure for efficiency coming from the cloud side. All right. Tim, before we get started, I think we should run a polling question.

Yes. Good time for our first poll. So our first question today is why do you care about Data Center energy efficiency? You can look through the answers and choose the best one for your needs, and we'll go ahead and show you the results here in a few slides.

Okay. Great. Yes, please go ahead and answer that polling question. And meanwhile, let's go on with content of the webinar. So I'm sure you're interested in how can I make my Data Center more efficient and reduce the cost. And there are really 5 ways that you can reduce the power costs in the data center. And let's talk about those. The first is, of course, you can find cheaper electricity. So maybe I'm using the same amount of electricity, but I move my data center to a place or I rent space in a data center that has cheaper electricity, might be hydroelectric, it could be wind, it could be solar, but that's definitely one way to reduce costs. It's not really making things more efficient, but it is lowering the price of the electricity. Second way is to improve power usage effectiveness. I'll talk more about that in just a minute. But this essentially means that a higher percentage of your electricity going into your data center is actually used to run your computing equipment, your servers, your networking and your storage. A third way is, outsource to the cloud. As I mentioned, in most cases, a cloud service provider has a data center that's already more power efficient than a typical enterprise or a private data center. So certainly, you can do that, but not all workloads can move to the cloud. There are definitely customers who are looking and saying, well, some workloads might go to the cloud, but some workloads I can run more efficiently with higher performance or at a lower cost in my own data center. And there may also be privacy or data governance or confidentiality reasons that you cannot move all your applications into the cloud. Now at the bottom of the slide we have 2, I think more interesting ways to improve -- reduce your power costs. One is to maximize server utilization, to make sure your servers are fully utilized and running as many applications as possible. And the other is to improve server power efficiency. So getting the same work done or more work done with less power consumption. And those 2 ways are interesting because that's where the DPU, or data processing unit, can provide a benefit to help achieve better power efficiency and lower your power costs. So we're going to focus on those 2 at the bottom, maximizing server utilization, and improving server power efficiency. Hey, Tim, do we have the results of that first poll?

Yes. Let's take a look real quick. So looks like taking the lead right now, power limitation and data center. So that was one of the items that you did mention at the very beginning. Save the world and environment comes in as a strong second. And mandate from top management doesn't seem to be getting anything in results there. But definitely saving money and requirements imposed by government regulators are up there as well.

Okay. Great. That's interesting. So it's actually I think gratifying to see that our audience out there is out to save the world and -- or maybe get more compute into their data centers and they're not doing this just because the boss said, hey, you have to be more efficient. But they may be -- a few of them may be doing it because the government said you have to be greener or more efficient. All right. So coming -- talking about that, why do we worry about this? And it turns out that power consumption in data centers has been a growing concern because there's pretty much an agreement that data centers are using about 1% of the world's electricity today or in 2020, about 1%. So of all the electricity produced, it's 1%. And actually, in the United States, it's actually higher. It's almost 2%, and it's almost 3% of electricity production in Europe. So it turns out these researchers did a very careful study. And in 2015, they looked and said, oh, we might be in trouble because data centers are set to use 8% of all global electricity by -- I'm sorry -- use 8% of global electricity by 2030 and then they revised -- in 2020, they went back and looked, okay, so it's not that bad, it's not going to be 8% or worst case, 13%. We now think it's going to be about 3% and 8% is now the worst case. But that's still a lot. So that's still -- to say that all the e-mails and the -- supporting your mobile phones and your web browsing, and e-commerce is going to consume a 3% or possibly 8% of all electricity in the world. So we really need to find a way to minimize that and try to keep it down more to the 1% or 2% range. And these researchers when they looked at it and when they updated their study in 2020, they said, we can hold down electricity consumption of data centers if, we keep improving the efficiency of those data centers and those servers every year. And they also assume that more of the workload is going to shift to the cloud because the cloud tends to be more efficient. So definitely, it's a concern but it turns out that, by improving efficiency, we can prevent data centers from eating all the world's electricity. Now in terms of the pricing, that's definitely a concern because I mentioned global electricity prices are going up. This is a chart of electricity prices in 3 European countries in the first half of 2022. You can see that they went up quite a bit for the -- for those in the U.S. and our audience, it turns out average electricity price in the U.S. tends to be a bit lower than in Europe, but we can see that not only are the prices higher in Europe, but they took a big spike in 2022 going up. Now since then, the prices have come down a little bit, but it's still higher, definitely higher than when they started in 2022. So if you look towards the end of '22 it turns out the winter was not as cold. They supplied -- got extra gas supply sent in by ship and the prices did moderate a bit. But definitely, that's still a significant concern because most people -- nobody really believes that those prices are going to come down back to where they were before. So I did mention, and when I said 5 ways to save on the cost of electricity for your data centers. One of them was improving PUE. So what is PUE? How can I improve it? And why is PUE itself not enough on the quest for a green or more sustainable data center? So it means power usage effectiveness, and it's simply the ratio of how much power is going into the data center versus how much is actually used by the servers, the networking, and the storage. So you may ask, "well, why would I need more electricity than what's actually used by the IT equipment?" And the answer -- one answer, there are really two answers to that. The first is that every time the electricity transitions, it goes through -- the voltage changes or it goes through UPS or goes into batteries or even it goes into the power supply of the computer, you lose some of that power, and it's turned into -- it's wasted or turned into heat. And second, you need -- for all the power in the data center, you need to use more electricity to cool the data center by extracting the heat and cooling it down. And then a little bit goes to like cooling -- sorry, lights and cameras and sensors and auto and door locks and things like that. And that's probably like 1% of the power. So it turns out that you can -- for every watt of electricity going into the data center, you might have 0.6 or 0.7 watts for cooling, depending on your environment and then some additional is lost in conversion. So electricity going to UPS, it changes the voltage, it may go from AC to DC and then back to AC, you lose some power. In the power supply of the computer, it goes from AC back to DC and you lose some of the electricity. So in any case, there are ways to make this more efficient, better power supplies, better transformers, better UPSs, LED lighting, hot and cold aisles. And all of that reduces the PUE, which means you get -- make your data center more efficient. But we see from this chart that PUE has already become more efficient over the last several years and it sort of plateaued. It's plateaued around 1.5, 1.6 in the average data center. And that's because as you get more efficient, it becomes harder and harder to lower the PUE even further. So for many data centers, getting below 1.5 is kind of difficult for big cloud data centers, actually, most of them have gotten down to 1.2 and are pushing it down to maybe even 1.15. So that's why there's an argument you can save electricity by moving to the cloud. So it doesn't always save you money for all workloads. So now that you've hit the limit, you can't improve PUE anymore. The next step is really to improve the efficiency of the servers. Tim, I think it's a good time to ask another polling question here with the audience.

Great. Okay. Next question is up on your screen. What is the current top priority to improve data center energy efficiency? Take a look at that and provide your answers. Real quick, John, we had a question came in. And we'll give the -- our attendees a chance -- to provide their answers, while we look at this question. And the question is DPUs is going to be much more to conserve energy. DPUs can -- so it's basically, it is the fastest, most power efficient, available, why are we not considering that? Pretty easy. We definitely are aware that DPUs are much more than just something to save power or improve power efficiency within the server. But that's a conversation for another topic. Today's focus is on power efficiency. Do you have anything to add to that, John?

No, Tim, I think that's a great answer. And I will talk about that a little bit when we explain how DPUs -- we're going to get into how DPUs can actually make the service more efficient. So okay, I'll go ahead and go forward, while we wait for the -- yes, we wait for the poll results. So okay. So let's talk about DPU? What is the DPU? Why are we talking about it? And how can it help with that power efficiency? I said it helps with what 2 factors, 2 of the 5 ways you can reduce your power costs. So DPU is a data processing unit. It's basically a chip that's specializes in offloading and accelerating that data center infrastructure, which means networking, security, storage, it could do encryption, it can do remote management. It can maybe -- you can do security tasks like running a firewall agent or screening, doing packet filtering. So it can do accelerated networking by accelerating networking for AI and for high-performance computing. So this is basically what a DPU specializes in. This is a picture of the NVIDIA BlueField-2 DPU, which is shipping right now and available through major server vendors. And then we're also sampling the BlueField-3 DPU, which is the next generation. So the key is that the DPU can do certain things more efficiently and faster and with less power than a CPU. Sorry, I think I jump the gun here. Tim, do we have the results from our last survey questions.

Let us take a look at this real quick. It looks like improving server efficiency is the #1 reason by far. It makes a lot a sense why people are joining the call today.

All right. Yes. And by the way, when I mentioned that outsourced to the cloud, and I've said that many cloud service providers are already more efficient. It turns out a lot of the big cloud service providers are already using DPUs, and that's part of how they achieve that higher power efficiency or better power efficiency in their data centers. All right. Let me go ahead and go forward. All right. So Tim, related to that question that came in, so yes, DPU not only can help save power, but it can accelerate -- it does certain things faster. And actually, the key is that, it typically does it more efficiently with less electricity, fewer watts per operation or more operations per watt. And so on the left, we see a typical setup without a DPU with a regular network card. If you have virtualization, you have overlay networks, you have software-defined security, software-defined storage. The CPU is having to do all of that. And up to 30% of the CPU's cores or attention is used to handle that infrastructure or those tasks, leaving only the 70% of the CPU to actually run applications. On the right side, if you have a DPU, the DPU is offloading and accelerating and in some cases, also providing performance and security isolation for these tasks. So now nearly 100% -- 95%, 100% of the CPU can be devoted to running the applications and the DPU is taking care of the networking, the storage, the security, remote management and monitoring and things like that. So functionally, from a functional standpoint, that's how the DPU is going to reduce power and make the servers more efficient. So I think now we should actually talk about some examples of what happens. So we thought this is very interesting. We had customers asking for this. So we actually did some testing. We said, let's go in and we'll test different applications, and we'll test them with a regular network card and with a BlueField DPU and see what happens. So we did some of this testing with partners and some on our own, and this is a testing that was done with Ericsson. Very well known in the telco world. Ericsson said, let's take a typical UPF, which is user plane function. It's a telco function, networking function. And it's actually -- even though it's telco-specific, the workload is actually very similar to many other software-defined networking workloads that an enterprise or a media company or a regular business might run a new data center. So think of this as a software-defined networking workload. This one just happened to be for telcos. And Ericsson said, "what if we -- well, let's look at a normal server, where with no power efficiencies. Let's look at -- in the blue bar, what if we let CPU go to sleep or slow down when it's not busy. And what if also put in a DPU, and we offload work or offload the networking to the DPU. And we see that when the server is not doing any work, the CPU power efficiencies saved a lot of power, the DPU doesn't really offer any additional benefit because there's no work going on to offload. At 50% load, we see that we save some -- a fair amount of power with CPU efficiencies and a little extra power with the DPU. But when the servers at 100% workload, you see that the CPU is not efficiencies, can't save so much power because the CPU is busy all the time. But the DPU is able to save a significant amount of power -- an additional 15% of power. And combined, letting the CPU take a break when it's not working and offloading networking into the DPU, we were able to save 23% or reduced power usage by 23%. So this is significant. And this was actually measured -- this was measured at the CPU. And it shows that if you have a big data center and you did this for 3 years because many servers live for 3 years, you would save $1.77 million over 3 years, just on the power from offloading to the DPU. So I think that's pretty significant. All right. We did another networking test for those who are doing software-defined firewalls, load balancers, just general software-defined networking then a very common use is something called OVS, and also now as Open vSwitch. And normally, this runs on the CPU as software define networking but you can offload it to the DPU and it's much more efficient, the DPU as silicon that's specialized in running OVS in a faster, more efficient way. So again, in this case, we measured the power at the server power supply. And we saw that the server consumed up to 30% -- 29% less power by offloading to the DPU. And again, you can see that the higher the workload on the server, the more power you save by offloading the network to the BlueField -- networking to the BlueField DPU. So in this -- if you -- on the left side, we look at the actual power savings. We see that the performance was 2.5x faster with the DPU and we saved 18 virtual cores. So that would be like 9 physical cores or 18 hyperthreaded cores. And that's quite a bit. That's 22% of the cores were released and freed up from doing networking and available to run the applications. So the power savings here -- just looking at the power savings, not looking at the likely ability to reduce the number of servers being used is 29%. And over 3 years, it's worth $5 million in a large data center. That's a lot of cash that you can pay for more IT administrators, more servers. So again, 3 years $5 million savings in a large data center just looking at the power savings per server. And again, not looking at the possibility, the fact that I can probably reduce the number of servers as well. All right. So what about security and encryption? So we did another test, and we said, okay, a lot of data centers with Zero-Trust, they want to encrypt everything because Zero-Trust says, I might not be able to trust all the servers in my data center, even inside the firewall perimeter. So I have to assume that any server could possibly get compromised. So one best practice in this case is to encrypt all the traffic between the servers. So this is typically done with IPsec, and what we did is we set up an IPsec server and an IPsec client, and we assume that we're writing. So in this case, we have one doing mostly encryption and one doing mostly decryption. And then actually, in this case, the server was doing -- the client was writing and doing most of the encryption, which is harder, and the server was receiving the data and mostly doing decryption. And again, we compared a regular NIC to the DPU and we measured the power consumption. And here, on the client, which is doing the encryption work, we saw a very large savings. So on the server, we saw 21% power savings. And we also saw faster performance -- I'm sorry, we saw 21% power savings, when the server was busy. There's really no power savings when the server is not doing anything, which makes sense. And on the client, which had the harder job of encrypting data versus decrypting, again, when it's idle, no difference, but when it's busy, there was a 34% power savings. So that's actually very significant. And you can see here that it's quite a big difference. And Actually, you know what, I think we have an error in these charts. So we'll actually look into that because these charts are the same. They should be slightly different. But in any case, the numbers on the left are correct, and it shows that on the server side, you would save $5.5 million a year. And on the client side, you'd save $8.7 million a year. And again, this is just looking at the power savings per server, and it's not accounting for the fact that they're -- we freed up 6 or 20 core server cores, CPU cores in these cases. So on the IPsec server, application server, 6 CPU cores, the physical cores were made available because the DPU had relieved it from having to do that encryption -- the decryption work. And on the client side, there were 20 cores. That's a huge amount. That's 20 out of 80 original cores were freed up and could be used to run applications or reduce the number of servers. And we'll talk more about that in just a second. So Finally, and we did many tests, but this is the final one we're presenting, but we present -- we tested a NoSQL database called Redis. It's used for caching and as a database or for caching in front of a database. And we ran on top of VMware vSphere because vSphere last August launch support for DPUs, including the BlueField DPU. So this allows VMware vSphere or EXS, ESXi to offload its networking to the DPU. And this allowed Redis to use fewer server -- to have more -- sorry, this freed up CPU cores, which allowed Redis to run faster. And we did a testing, and this is the most conservative test. We're actually doing more of this testing even today and finding better performance improvements and good power savings. Where you can see that there are 12 cores that freed up that are no longer needed to run the VMware networking. So on the left side, you can see that with a regular NIC, 12 core is used for networking with a DPU, 0 core is used for networking because the CPU is free to do things like Redis or other administrative or application workloads. And the performance is slightly higher. And as you increase the number of Redis instances on the server. So you get slightly better performance and of lot many fewer cores. So based on this, we calculated -- you could reduce the number of servers by 15%. And because of the software licensing and the physical costs and the power costs, you save money on the number of servers, you reduce your power consumption. And you would say $56 million. In this case, we had a large data center, running a lot of Redis and you introduce DPUs, you get a huge savings of More than $50 million over 3 years. So that is actually a very large amount. And again, it's all because the DPU can do the networking, storage and encryption and other security tasks more efficiently than a CPU. And we're actually doing more of this testing with Redis and VMware right now. We hope to publish the results soon. And we're also working on more integration. So for example, there will be other VMware features that will be offloaded -- that are not offloaded to the DPU today but will be later in this year. So even more efficiency is on the way for our VMware customers. And we're doing similar testing with Red Hat, with [indiscernible] And so there are people who are running, for example, Linux and containers will also be able to see power savings and better performance by offloading that infrastructure work to the DPUs. All right. I think this is getting towards the end of our TCO and power savings examples. So earlier, I said, we have testing where we only calculated the savings by reducing the power in each server. And we did not -- we thought there would be savings from reducing the number of servers, but we haven't calculated those, except for the Redis and VMware example I just showed. So we actually went back to the encryption example with IPsec. And we said, while we know we're getting better performance and we're using fewer CPU cores by offloading to the DPU. This means we wouldn't need as many servers. So how much power could we say if we -- each server uses less power and we don't, we need fewer servers. So this is all detailed here in this chart. I know it's a lot of numbers, but it simply says that, hey, if I take -- use a regular NIC versus using a DPU, I start with 10,000 servers. With the offload I have to the DPU, you have to remember that we were able to save 23% or 34% of the power. And in addition, we were able to reduce either save 6 or 20 CPU cores depending on the encryption workload. So in this case, we found that we could reduce the number of servers by 18%. The cost per server goes up slightly because the DPU costs more than a regular NIC. So I'm paying a little more per server, but I have 18% fewer servers. So this means that I save $6.4 million or 6.3% on the CapEx, capital expenditure, the cost of buying the servers. The power reduction is per server is 34%, as we showed before. So the total power of savings over 3 years is 45% because Tim, we've got fewer -- 18% fewer servers, 34% less power per server, and that's a total of 45% -- almost 46% reduction in total power. And that's worth $13 million at a typical California power pricing, $0.15 per kilowatt hour. So if you were in the U.K., if you were in Spain, your savings will be significantly bigger because the power costs more there. Now if we look at the power, we assume the power usage effectiveness of 1.5, and that means that the actual total savings -- the savings for the server power consumption is $13 million, but the data center is saving almost $20 million. It's $19.7 million because with a PUE of 1.5, I probably should have explained this earlier. It means that every watt of power that a server needs or that network switch needs requires 1.5 watts of power to go into the data center. So likewise, it means, if I save, $13 million in the server power consumption, I actually save $19.7 million in the data centers power consumption because, again, every watt of power used by the IT equipment requires 1.5 watts of power into the data center. And of course, the utility charges you for every watt going into the data center, not just for the watts that are used by the servers. So the bottom line, literally the bottom line on this chart is that for a large data center that needs to encrypt everything and is running applications on top of that, by going to the DPUs, you would save $26 million, and that was what we highlighted early on at the beginning of this webinar that -- and that's a 17.8% TCO savings. Again, over 3 years, assuming a typical PUE and typical California or typically U.S. power pricing. So that's really, I think, very big. $26 million is very large savings. And we say it's a combination of CapEx and operating expenses. And the bigger part of the savings is, in fact, on the operating expenses and the reduction in power consumption on top of the reduced cost of the equipment because I need fewer servers. All right. So Tim, where are we now? I believe, yes. So I think some people may be asking, hey, this sounds interesting, but do any application support the DPU? Or what infrastructure, what type of container management or hypervisor can I use with the DPU? Do any of my applications work with it? So the good news is that the BlueField DPU has an ecosystem. There's a development platform called DOCA, which makes it very easy to integrate applications with the DPU or to develop new applications that take advantage of the DPUs, acceleration and offloads. And this is a partial list of some of the partners that support the BlueField DPU today. So for example, on the infrastructure side, we talked about the testing we did with VMware. So obviously, they had their big launch in August of last year. We've published a blog and paper with Red Hat, and we're doing more testing with them, where Red Hat OpenShift can offload to the DPU to gain that efficiency and save power. And we're also working with Canonical. And then if you look at cybersecurity, we have several partners. For example, Palo Alto Networks, that's available today, where they offload to the DPU and they get more power, they get -- sorry, faster performance. I think it's a 3 to 5x faster networking performance for their distributed firewall, but they also free up CPU cores. You free up 5 or 6 CPU cores, which can be used for other things. And that itself is either going to let you run more applications or it's going to reduce power consumption on the server or both. And then we have additional firewall and security companies that have integrated with the BlueField DPU. We also have, for example, Juniper it's called JESP. I forgot exactly JESP stands for. I think it's a Juniper Edge Services Platform. So Juniper JESP is integrated with DPU. F5 has a product that's also integrated with BlueField as Broadridge Networks. And then these are 3 storage vendors who are supporting the DPU now to offload storage functionality and networking and make their storage products faster and more efficient. And definitely, efficiency is part of the reason why these storage vendors are using the DPU. So I think the rest of the details just to give a high an overview of how DOCA works with the DPU. It also means that people who develop applications on one version of the BlueField will have it supported on the next version, and the next version and so forth. Right, Tim, I think we can...

Are we ready for a few questions?

Yes. We are...

Some of these were -- some of these questions were covered a little bit, but maybe we could provide a little bit more detail. They might have been asked, some of them I think might have been asked before you covered the topics. But we'll just jump in and see what we can do to answer these. We had a question, someone was asking, what type of encryption was used, was it AES or RSA? I don't know if that was listed in there.

It was not listed. But in this test with IPsec, it was AES-256. The BlueField supports offload of multiple types of encryption. And I think we added even more types of encryption with BlueField-3. But in this case, it was AES. And the BlueField-2 can do that, use that encryption for IPsec, which is good for typical server-to-server communication in a data center. They can do it for TLS, which is good for web traffic that TLS is typically -- it's the successor to SSL. And so that's for web applications, that would probably be the encryption of choice. I can also do it for storage encryption as well for Data At Rest Encryption of the storage.

Great. Next question comes in and they're asking about can users program DPUs with additional programming models? And they're asked specifically about [ MPI ] or OpenSHMEM. Do you want to get into a little bit on the DOCA drivers and DOCA libraries that are available?

Sure. I don't think the DOCA slide I used shows the details. But yes, you can program the DPU to do MPI offloads. And actually, the DPU can support that with ethernet or with InfiniBand. It actually, the MPI offloads are typically with InfiniBand, but the DPU supports both. That's very popular with high-performance computing and some types of AI workloads as well. What was -- Tim, what was the other question about the programming. It was...

It was OpenSHMEM and NVSHMEM?

I'm pretty -- I'm not familiar with NVSHMEM, but I believe the DPU can be programmed to support that as well. So DOCA, we have some web pages on our -- it's not listed here in the resources page, but we have some -- plenty of web pages on DOCA on the NVIDIA website that have diagrams and talk about what's supported. And there's a whole road map of DOCA features in. And that I believe those are included either supported today or very soon in the DOCA framework.

Great. Thanks, John. Next up -- our next question comes in. Do you have an AI or scientific computing application results for those?

We do have performance results. We have several, but we do -- we do not have direct power measurement results, but what we do have for scientific computing -- okay, yet we will be doing power measurements with AI and scientific computing workloads. What we do have is, in November at Supercomputing, we published as it happens every 6 months. Of course, there was a list of the Top 500 and the Green500, most powerful supercomputers and the most efficient supercomputers in the world. And it turns out that some of the most efficient computers are using both NVIDIA GPUs and NVIDIA Networking and with either InfiniBand or Ethernet. And some of them are using the DPUs. So a lot of these offloads we talked about can -- have been proven to show either with our ConnectX or our BlueField DPU have been shown to increase the efficiency of supercomputers, which are typically used for scientific computing and AI. Now we do expect to have some direct tests of AI workloads where we're going to run AI workload with them, without the DPU and measure the power consumption. Don't have that in the white paper or in the Resources section today, but it is coming.

Great. Next question here. I don't know if we've done any testing with KVM, but it's basically what we expect these results to be hypervisor agnostic.

Most likely, yes. Though -- so with KVM, I think -- I think today, with KVM, you would save power with the DPU and you would get better performance, but it might not -- the savings might not be as big as you have today with VMware, simply because VMware has done some specific integrations with the BlueField DPU and to offload more parts of the hypervisor and all the networking to the DPU. I'm not sure we've got that the same level of integration with KVM yet. So I'm going to say with KVM, you have power savings, but they might not be -- the power savings may not be as big until we get tighter integration with KVM.

An interesting question here. It's just do you have estimates on engineering time required to port application DPUs. I think it's a pretty broad question because it may depend on the engineering experience, it may depend on your experience with DOCA, the drivers. What you're trying to port, whether we have DOCA drivers and we have DOCA libraries. And the DOCA driver, you could basically it's just kind of out-of-the-box experience where you can get really good performance, wherever it's optimized. And the DOCA libraries tends to be a little bit more engineering specific where you need to put more time into that. So I don't know if we can estimate that. It may depend on the engineering experience that they have. And if we have the libraries that will support that are trending -- the application you're trying port.

Yes, Tim, I think your answer is spot on. If they have an application that already is using Open vSwitch or already using DPDK or SPDK which are supported inside DOCA, then porting could be very fast. An application that's using those, and you just want to port that part to the DPU, that part of the networking of storage. Okay. As you said, it depends on the application and the experience of the engineers I'm to say, okay, 1 week for those type of applications. Now if it's an application that's going to do a more sophisticated integration like on the level of what VMware did to support the DPU, then it's definitely going to take more time. But the good news is if you're running on top of VMware or running on top of OpenShift, they already did the work. So you can get a lot of those benefits with no work -- no porting if you just run your application on top of VMware or OpenShift. But unlike Palo Alto Networks, okay, they took more than a week. I don't know exactly how long it was, but they actually moved -- they have some called intelligent traffic offload, where the traffic comes in and they quickly classify, is this traffic that requires more personalized inspection or traffic they know how to deal with. If they know how to deal with it, they offload all that to the DPU and they get better performance and reduce the number of CPU cores needed. At this traffic they haven't seen before or require special attention or suspicious then, they process it normally through their -- the CPU. So that's why they call it intelligent traffic offload. And that took more than a week, but it's not that difficult, especially now that we have -- as we said, the DOCA drivers and libraries and DOCA include sample applications as well that can make it easier. So I guess the short answer is, work required -- a very small amount of work if you're using a library that's already in DOCA, like DPDK or OBS. And more work if you're not using one of those libraries.

Good answer, John. It's kind of difficult one to get to. And here's another difficult question as well, but I think we can give a general answer to it as well. I'm looking to see...

Sorry, Tim. Just want to mention the resources here. So just want to make sure that, if anyone is interested in more information about the DPU, there's a white paper, there's a blog, there's information about the BlueField. We have some of -- there's a blog about Red Hat OpenShift down Bluefield. There's information on VMware on Bluefield. And there's a lot of information about DOCA. So you can all -- you can contact us through -- there is a link here, a web form. If you go on the Resources Tab along '24 of this webcast and then you'll see a lot of links to these resources. And we would love to have you fill out a survey at the end of the webinar to tell us what you thought. All right, Tim. Sorry for that little commercial about the slide. But please go ahead.

Now we had to squeeze it in somewhere. Good time. Okay. So this one is kind of the question is, is there a way that they can kind of figure out what the power savings might be from DPU or is there a finite number of watts per DPU that can be applied towards savings? And I think the broad question because we've seen a lot of different results, but we are seeing power savings in everything. I might throw out 30%, but let's just what's your -- you might suggest, John.